fix(a2a): return long_running_tool_ids from the message and artifact update converters

[Yatsuiii]

_convert_a2a_parts_to_adk_parts recovers long-running function call ids from
the is_long_running part marker for all four inbound converters. Two of them
discard the result before building the Event.

converter	passes ids to _create_event
convert_a2a_task_to_event	yes
convert_a2a_status_update_to_event	yes
convert_a2a_message_to_event	no
convert_a2a_artifact_update_to_event	no

Both failing sites now keep the ids and pass them positionally, matching the
two that already worked.

Behaviour change

Both are defaults on A2aRemoteAgentConfig. On the new integration extension
path _should_pause_invocation saw an empty long_running_tool_ids and did
not pause, so a long-running call ran on as if it had completed. It now pauses,
matching the legacy event_converter path.

Tests

One parametrized case across all four converters. Without the source change it
fails on message and artifact_update and passes on the other two.
tests/unittests/a2a is green: 460 passed, 48 skipped.

Fixes #6988

[mahirhir]

The same discard is still on line 510, in convert_a2a_task_to_event, which is one of the four converters your test parametrises:

output_parts, _ = _convert_a2a_parts_to_adk_parts(
    artifact_parts, part_converter
)

The task fixture in _LONG_RUNNING_INBOUND_CONVERTERS builds a Task with a status.message and no artifacts, so it takes the status branch on line 523, which was already keeping the ids before this PR. A Task that carries the long-running call in an artifact never reaches that branch, and that is the same shape you just fixed in convert_a2a_artifact_update_to_event.

There is a second reason to cover it. _create_mock_function_call_for_required_user_input returns early only when long_running_function_ids is non-empty. With the ids dropped on 510 the set is empty, so an input-required task that already carried a genuine long-running call falls into the synthesis path instead of being left alone. If any converted part has text, that part is replaced by a synthetic call under a fresh uuid, and the caller is then holding an id to answer that is not the id of the call that is actually pending. Losing the id and inventing a different one are not the same failure, and the second one is quieter.

The repair looks like the one you already wrote twice: bind the ids at 510 and long_running_function_ids.update(ids) before the status branch adds its own.

[Yatsuiii]

Good catch, thanks. Confirmed both parts.

Line 510 was a third discard site, and my parametrised task case never reached it. The fixture builds a Task with a status message and no artifacts, so it took the status branch that was already keeping the ids. The case passed without covering the path it was named for.

The synthesis point holds too. With the ids dropped there, an input-required task falls past the early return in _create_mock_function_call_for_required_user_input, and a genuine pending call is replaced by one under a fresh uuid. Quieter than losing the id outright, as you said.

Fixed in 2a32cc9, same shape as the other three sites. Added a task_artifact case that routes a Task through the artifact branch, plus a test pinning that a real call is not replaced. Both fail without the change; tests/unittests/a2a is green at 462 passed.

@llalitkumarrr flagging that the commit you tested was missing this third site, so it is worth re-running before review.